Automatic completion of underground utility topologies using graph convolutional networks

The absence of utility data, particularly about topological information, presents a significant impediment to the efficient management of underground utilities. Previous studies predominantly focus on general attributes such as diameter and material missing, neglecting the imperative issue of insufficient topological information. To address this gap, this study proposes the underground utilities topology completion (UUTC) model based on graph convolutional network (GCN) techniques. A comprehensive evaluation of the proposed model was conducted by performing completion experiments on a real public wastewater network database. This evaluation employed five prominent GCN models while simulating varying missing rates of topological data. The empirical findings indicate that the UUTC model exhibits a substantial advantage over the baseline models, achieving an average completion accuracy of 85.33%. The findings hold the potential to significantly mitigate the expenses associated with manual inspections from incomplete databases.